CT图像纹理分析鉴别乏血供胰腺神经内分泌肿瘤与胰腺导管腺癌

The texture analysis of CT images used for the discrimination of nonhypervascular pancreatic neuroendocrine tumors from pancreatic ductal adenocarcinomas

目的探讨采用CT图像纹理分析鉴别乏血供胰腺神经内分泌肿瘤(pancreatic neuroendocrine tumors,PNET)和胰腺导管腺癌(pancreatic ductal adenocarcinomas,PDAC)的可行性。方法回顾性分析四川大学华西医院于2009年1月至2017年1月期间收治的经病理学检查证实的乏血供PNET(共计15个病灶)和PDAC(共30个病灶)的CT资料。结果利用Ma Zda软件中的费希尔参数法(Fisher)+最小分类误差与最小平均相关系数法(PA)+相关信息测度法(MI)联合法自动选择出30个最佳纹理特征,在动脉期的频率分布为:共生矩阵18个,游程矩阵10个,自回归模型2个;在门静脉期的频率分布为:共生矩阵15个,游程矩阵10个,灰度直方图1个,绝对梯度1个,自回归模型3个。在动脉期和门静脉期中,Teta2均为诊断效能最高的单个纹理特征,其曲线下面积(AUC)值分别为0.829和0.740(P<0.001,P=0.009)。利用Ma Zda自带的B11数据分析模块分析30个最佳纹理特征,在动脉期,原始数据分析(RDA)/K邻近分类(KNN)法、主成分分析(PCA)/KNN法、线性判别分析(LDA)/KNN法和非线性判别分析(NDA)/人工神经网络(ANN)法的错判率分别为28.89%(13/45)、28.89%(13/45)、0(0/45)及4.44%(2/45);在门静脉期,RDA/KNN、PCA/KNN、LDA/KNN及NDA/ANN法的错判率分别为35.56%(16/45)、33.33%(15/45)、4.44%(2/45)及11.11%(5/45)。结论 CT图像纹理分析鉴别乏血供PNET与PDAC是可行的,其中纹理特征"Teta2"具有较高的诊断效能,动脉期LDA/KNN法具有最小的错判率。

Objective To determine feasibility of texture analysis of CT images for the discrimination of nonhypervascular pancreatic neuroendocrine tumor （PNET） from pancreatic ductal adenocarcinoma （PDAC）. Methods CT images of 15 pathologically proved as PNETs and 30 PDACs in West China Hospital of Sichuan University from January 2009 to January 2017 were retrospectively analyzed. Results Thirty best texture parameters were automatically selected by the combination of Fisher coefficient （Fisher）＋classification error probability combined with average correlation coefficients （PA）＋mutual information （MI）. The 30 texture parameters of arterial phase （AP） CT images were distributed in co-occurrence matrix （18 parameters）, run-length matrix （10 parameters）, and autoregressive model （2 parameters）. The distribution of parameters in portal venous phase （PVP） were co-occurrence matrix （15 parameters）, run-length matrix （10 parameters）, histogram （1 parameter）, absolute gradient （1 parameter）, and autoregressive model （3 parameters）. In AP and PVP, the parameter with the highest diagnostic performance were both Teta2, and the area under curve （AUC） value was 0.829 and 0.740 （P〈0.001,P=0.009）, respectively. By the B11 of MaZda, the misclassification rate of raw data analysis （RDA）/K nearest neighbor classification （KNN）, principal component analysis （PCA）/KNN, linear discriminant analysis （LDA）/KNN, and nonlinear discriminant analysis （NDA）/artificial neural network （ANN） was 28.89% （13/45）, 28.89% （13/45）, 0 （0/45）, and 4.44% （2/45）, respectively. In PVP, the misclassification rate of RDA/KNN, PCA/KNN, LDA/KNN, and NDA/ANN was 35.56% （16/45）, 33.33% （15/45）, 4.44% （2/45）, and 11.11% （5/45）, respectively. Conclusions CT texture analysis is feasible in the discrimination of nonhypervascular PNET and PDAC. Teta2 is the parameter with the highest diagnostic performance, and in AP, LDA/KNN modality has the lowest misclassification rate.